SKY API ChatGPT Bot Integration?

Christopher Hotchkiss
Christopher Hotchkiss New Member
First Anniversary
in SKY Developer General Discussions

I'm following a bit of a rabbit-hole, and wanted to see if anyone has tried to integrate a chatbot, via ChatGPT, into your NXT system via SKY API to serve as a sort of ‘Data Assistant?’ I'm researching, and it sounds like it's possible (allowing for data retrievals, updates, metrics, etc), but I wanted to see if anyone has had any success with such a thing.

Tagged:

Comments

  • Matt Thacker
    Matt Thacker New Member
    Fourth Anniversary Kudos 3 Name Dropper Participant

    Hey @Christopher Hotchkiss, I've not integrated a chat assistant directly into NXT itself, but I did build a command line application where I can chat or ask questions for my Blackbaud data. I imagine it'd be fairly simple to take the logic and build as a SPA within NXT.

    The CLI I wrote runs a couple of scripts. The first does a nightly pull of NXT data and writes the new and updated data to a Postgres db. Another script that performs some transformations to denormalize the data and prep it for embedding. Here, you'd want to use a library like LlamaIndex or LangChain. These will allow you to use different LLM models to perform different actions within the pipeline. A lot of applications are just a wrapper on top of the OpenAI gpt API, but that can get wildly expensive if you're using the OpenAI API for embeddings and retrievals. Based on how you format your data, I'd suggest trying different models for your embeddings and see what gives you the most accurate vectors. Store those embeddings in a vector db, like Chroma or Pinecone. The last step is to create the RAG atop the vector db.

    Honestly, the application isn't difficult to set up; it's understanding your particular use case. For example, the application I built was for my own use. I built everything using models I'm running locally on my computer via Ollama and using a locally stored instance of ChromaDB for my embeddings so the application doesn't cost me anything to run. The downside is, no one can run it. The code itself is reusable, but they'd need to have the same set up, whereas if you build using LLM API's, it's more deployable, but there are costs associated with every LLM.


  • Glen Hutson
    Glen Hutson Blackbaud Employee
    Tenth Anniversary Kudos 5 First Reply Name Dropper

    @Christopher Hotchkiss
    Great to hear the thinking about combining these items together. One thing to note is that Blackbaud is developing a native Copilot to be a Data Assistant. See https://www.blackbaud.com/newsroom/article/blackbaud-showcases-the-future-of-ai-powered-fundraising-and-financial-management-for-social-impact-organizations-at-bbcon-2024 . Key thing about this is that it will be built with data safeguards. Not to deter you from your own research and investigations, but thought to mention just in case.

Categories